AI-Enabled System Design — Agent System

New round introduced April 16, 2026 in the AI-Native full loop. Design an LLM-agent system for an automation task (auto-ticket-triage and code-review agent are confirmed prompts). ReAct-loop vs multi-agent decomposition + cost / memory / tool design are the live drill-downs. Requirements Confirmed prompts so far: (a) ticket-triage agent — read alert, gather context, propose resolution; (b) code-review agent — review a PR diff and produce structured comments. Reported deep-dive axes (every reported loop hit several): Agent shape: single ReAct loop vs multi-agent pipeline (e.g. triage → investigate → resolve). Interviewer may push you off your first choice. Which parts need an LLM vs deterministic code. The candidate is expected to justify each LLM call rather than wrap everything in one. Tool catalog: how tools are registered, described, rate-limited, authorized. Memory: short-term scratchpad vs long-term vector store; eviction policy. Orchestrator design: planning loop, retries, max-iteration cap. Validation of agent outputs: golden-set replay, structured validators, human-in-the-loop on write actions. Cost accounting: per-call token cost, expected cost per ticket / PR; budget cap and degradation strategy. Operational concerns: hallucination detection, idempotent tool retries, observability. Notes One candidate reports the interviewer steering them from a triage→investigate→resolve multi-agent design toward a ReAct loop. Confirm topology preference before committing. Another candidate's loop reported "AI assist" listed in the prep materials but absent in the actual interview environment — interviewer said "move on without it." Be ready to whiteboard the design without live AI hands-on. Sister round to traditional System Design — same 45-min budget, agent-specific evaluation rubric. Coverage is still thin; expect prompt mutation through Q3 2026. Preparation Study published agent designs: ReAct, Reflexion, AutoGen multi-agent, Anthropic's Claude tool-use patterns. Pre-write a single-page template: tools, memory, planning loop, validators, cost, guardrails. Memorize a per-LLM-call cost model: tokens-in × tokens-out × rate; back-of-envelope cost per ticket. Rehearse defending both single-agent ReAct AND multi-agent pipeline for the same prompt. Re-check the forum 1-2 weeks pre-interview; this round is new and detail keeps surfacing.